README.md

iMaterialist Challenge Furniture 2018 - Playground

Example of model training/validation/predictions on iMaterialist Challenge Furniture 2018 dataset.

/!\ Python3 only code

Download data from Kaggle

kaggle competitions download -c imaterialist-challenge-furniture-2018

and download images to your local storage:

python utils/download_images.py train.json input/train
python utils/download_images.py validation.json input/validation
python utils/download_images.py test.json input/test

Total dataset size is ~ 110GB

Resize datasets

python utils/resize_dataset.py input/train 224 input/train_224x224

Training a single model

Edit a configuration file, for example train_vgg16_bn_basic.py and simply run

python train.py configs/your_config.py

Predictions with a single model

Edit a configuration file, for example predict_vgg16_bn_basic.py and simply run

python predict.py configs/your_config.py

Simple blending

Dataset is splitted already on two splits: train and validation. Complete procedure is the following:

• Fit and predict with networks:

  • Fit network 1 on the train dataset
  • Predict probabilites with the network 1 on the validation and test datasets
  • Fit network 2 on the train dataset
  • Predict probabilites with the network 2 on the validation and test dataset

  ...

  • Fit network n on the train dataset
  • Predict probabilites with the network n on the validation and test dataset

• Create probabilites dataset composed of predictions on the validation dataset

  • Concatenate predictions of single networks

• Create probabilites dataset composed of predictions on the test dataset

  • Concatenate predictions of single networks

• Fit a meta-model on the validation probabilites dataset

  • Use cross-validation to estimate the performance and tune hyperparameters
  • Finally, train on the whole dataset

• Predict classes with trained meta-model on the test probabilites dataset